The Convergence Rate of Accelerated Proximal Gradient Methods with General Inertia Term for Convex Composite Multiobjective Optimization
Abstract
We study accelerated proximal gradient methods with general stepsizes for convex composite multiobjective optimization, in which every objective is the sum of a smooth convex term and a proper closed convex extended-valued term, and the proximal subproblem retains a separate offset for each objective. The momentum coefficient may depend on the computed trajectory. For a two-parameter root momentum rule we prove boundedness, weighted square summability of the displacements, a merit-function rate of o(k−2), and displacement decay of o(k−1) whenever the first parameter converges to a value in (0,1); the second parameter need not converge or have bounded total variation. The analysis is organized as a chain of basic estimates, an energy identity, coefficient asymptotics, and auxiliary lemmas that are reused in every regime. For the critical parameter α=3 we establish the finite bound u0(xN)≤ 2ℓ R/[N(N+3)], square summability of the displacements, displacement decay O(1/N), and convergence of the full sequence to a weak Pareto point. The remaining little-o question at α=3 is shown to be exactly the vanishing of a nonnegative critical energy, so it is separated from the large-O estimate rather than left as a missing step. For 1<α<3 we give the rate O(N−(α−1)) and for every α>0 the qualitative conclusion u0(xk)→0, Fi(xk)→ ci, and ∑k Dk/k<∞. A table summarizes the rate picture over the whole positive range of α, and a JOS1 experiment illustrates the predicted little-o separation.
Related
- Accelerated Proximal Gradient Method for Non-Strongly Convex Time-Varying Optimization
- A Safeguarded Accelerated Proximal ADMM Algorithm for Solving Structured Cardinality Penalized Optimization Problems
- Linear convergence rate analysis of Proximal-Tracking for nonsmooth distributed optimization
- Relative-Error Composite Federated Optimization with One Server-Side Proximal Step per Round
- Complexities of Weak Proximal Oracle Methods for Composite Convex Optimization
BibTeX
@misc{huang2026convergence,
title = {{The Convergence Rate of Accelerated Proximal Gradient Methods with General Inertia Term for Convex Composite Multiobjective Optimization}},
author = {Chengzhi Huang},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.06125},
eprint = {2611.06125},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.06125},
}
RIS
TY - UNPB
TI - The Convergence Rate of Accelerated Proximal Gradient Methods with General Inertia Term for Convex Composite Multiobjective Optimization
AU - Chengzhi Huang
PY - 2026
DA - 2026/11/04
PB - arXiv
DO - 10.48550/arxiv.2611.06125
UR - https://arxiv.org/abs/2611.06125
AB - We study accelerated proximal gradient methods with general stepsizes for convex composite multiobjective optimization, in which every objective is the sum of a smooth convex term and a proper closed convex extended-valued term, and the proximal subproblem retains a separate offset for each objective. The momentum coefficient may depend on the computed trajectory. For a two-parameter root momentum rule we prove boundedness, weighted square summability of the displacements, a merit-function rate of o(k^−2), and displacement decay of o(k^−1) whenever the first parameter converges to a value in (0,1); the second parameter need not converge or have bounded total variation. The analysis is organized as a chain of basic estimates, an energy identity, coefficient asymptotics, and auxiliary lemmas that are reused in every regime. For the critical parameter α=3 we establish the finite bound u_0(x^N)≤ 2ℓ R/[N(N+3)], square summability of the displacements, displacement decay O(1/N), and convergence of the full sequence to a weak Pareto point. The remaining little-o question at α=3 is shown to be exactly the vanishing of a nonnegative critical energy, so it is separated from the large-O estimate rather than left as a missing step. For 1<α<3 we give the rate O(N^−(α−1)) and for every α>0 the qualitative conclusion u_0(x^k)→0, F_i(x^k)→ c_i, and ∑_k D_k/k<∞. A table summarizes the rate picture over the whole positive range of α, and a JOS1 experiment illustrates the predicted little-o separation.
ER -